-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_generator.py
362 lines (297 loc) · 17.2 KB
/
train_generator.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
import argparse
import logging.handlers
import os
import random
from collections import OrderedDict
from torch.optim.lr_scheduler import MultiStepLR
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
from MultiWOZ import get_batch
from evaluator import evaluateModel
from tools import *
from transformer.Transformer import RespGenerator, UncertaintyLoss
def parse_opt():
parser = argparse.ArgumentParser()
parser.add_argument('--option', type=str, default="train", help="whether to train or test the model", choices=['train', 'test', 'postprocess'])
parser.add_argument('--emb_dim', type=int, default=128, help="the embedding dimension")
parser.add_argument('--dropout', type=float, default=0.2, help="dropout rate")
parser.add_argument('--resume', action='store_true', default=False, help="whether to resume previous run")
parser.add_argument('--batch_size', type=int, default=3, help="train/dev/test batch size")
parser.add_argument('--model', type=str, default="model", help="path to save or load models")
parser.add_argument('--data_dir', type=str, default='data', help="data dir")
parser.add_argument('--beam_size', type=int, default=2, help="beam size of act/response generator")
parser.add_argument('--max_seq_length', type=int, default=50, help="max input length")
parser.add_argument('--ngram', type=int, default=3, help="avoid n gram repeatness")
parser.add_argument('--layer_num', type=int, default=3, help="transformer layer num")
parser.add_argument('--evaluate_every', type=int, default=5, help="checkpoints")
parser.add_argument('--head', type=int, default=4, help="head num for transformer")
parser.add_argument("--learning_rate", default=1e-3, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--output_file", default='output', type=str, help="path to save generated act/response")
parser.add_argument("--non_delex", default=False, action="store_true", help="non delex testing")
parser.add_argument("--hist_num", default=0,type=int, help="turn num of history")
parser.add_argument('--log', type=str, default='log', help="log file")
parser.add_argument('--act_source', type=str, choices=["pred", "bert",'groundtruth'], default='pred', help="action source for validate/test")
parser.add_argument('--seed', type=int, default=1, help="random seed for initialization")
args = parser.parse_args()
return args
args = parse_opt()
if args.option == 'train':
if not os.path.exists(args.model):
os.makedirs(args.model)
args.log = os.path.join(args.model, 'train.log')
elif args.option == 'test':
dir = os.path.dirname(args.model)
args.log = os.path.join(dir, 'test.log')
logger = logging.getLogger(__name__)
handler1 = logging.StreamHandler()
handler2 = logging.FileHandler(filename=args.log)
logger.setLevel(logging.DEBUG)
handler1.setLevel(logging.WARNING)
handler2.setLevel(logging.DEBUG)
formatter = logging.Formatter("%(asctime)s %(name)s %(levelname)s %(message)s")
handler1.setFormatter(formatter)
handler2.setFormatter(formatter)
logger.addHandler(handler1)
logger.addHandler(handler2)
device = torch.device('cuda' if torch.cuda.is_available() else "cpu")
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
numpy.random.seed(seed)
random.seed(seed)
setup_seed(args.seed)
with open("{}/vocab.json".format(args.data_dir), 'r') as f:
vocabulary = json.load(f)
act_ontology = Constants.act_ontology
vocab, ivocab = vocabulary['vocab'], vocabulary['rev']
tokenizer = Tokenizer(vocab, ivocab, False)
with open("{}/act_vocab.json".format(args.data_dir), 'r') as f:
act_vocabulary = json.load(f)
act_vocab, act_ivocab = act_vocabulary['vocab'], act_vocabulary['rev']
act_tokenizer = Tokenizer(act_vocab, act_ivocab, False)
logger.info("Loading Vocabulary of {} size".format(tokenizer.vocab_len))
# Loading the dataset
checkpoint_file = args.model
if 'train' in args.option:
*train_examples, _ = get_batch(args.data_dir, 'train', tokenizer, act_tokenizer, args.max_seq_length)
train_data = TensorDataset(*train_examples)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.batch_size)
*val_examples, val_id = get_batch(args.data_dir, 'val', tokenizer, act_tokenizer, args.max_seq_length)
dialogs = json.load(open('{}/val.json'.format(args.data_dir)))
gt_turns = json.load(open('{}/val_reference.json'.format(args.data_dir)))
elif 'test' in args.option or 'postprocess' in args.option:
*val_examples, val_id = get_batch(args.data_dir, 'test', tokenizer, act_tokenizer, args.max_seq_length)
dialogs = json.load(open('{}/test.json'.format(args.data_dir)))
if args.non_delex:
gt_turns = json.load(open('{}/test_reference_nondelex.json'.format(args.data_dir)))
else:
gt_turns = json.load(open('{}/test_reference.json'.format(args.data_dir)))
eval_data = TensorDataset(*val_examples)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)
BLEU_calc = BLEUScorer()
F1_calc = F1Scorer()
best_BLEU = 0
weighted_loss_func = UncertaintyLoss(2)
weighted_loss_func.to(device)
resp_generator = RespGenerator(vocab_size=tokenizer.vocab_len,
act_vocab_size=act_tokenizer.vocab_len,
d_word_vec=args.emb_dim,
act_dim=Constants.act_len,
n_layers=args.layer_num,
d_model=args.emb_dim,
n_head=args.head,
dropout=args.dropout)
resp_generator.to(device)
bce_loss_func = torch.nn.BCEWithLogitsLoss()
bce_loss_func.to(device)
ce_loss_func = torch.nn.CrossEntropyLoss(ignore_index=Constants.PAD)
ce_loss_func.to(device)
label_list = Constants.functions + Constants.arguments
if args.option == 'train':
resp_generator.train()
if args.resume:
logger.info("Reloaing the encoder and act_generator from {}".format(checkpoint_file))
logger.info("Start Training with {} batches".format(len(train_dataloader)))
optimizer = torch.optim.Adam(filter(lambda x: x.requires_grad, list(resp_generator.parameters()) + list(weighted_loss_func.parameters())), betas=(0.9, 0.98), eps=1e-09)
scheduler = MultiStepLR(optimizer, milestones=[50, 100, 150, 200], gamma=0.5)
alpha = 0.1
for epoch in range(51):
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, action_masks, rep_in, resp_out, belief_state,\
bert_act_seq, act_in, act_out, all_label, act_input_mask,\
resp_input_mask, *_ = batch
resp_generator.zero_grad()
# act loss
logits, _, act_vecs = resp_generator.act_forward(
tgt_seq=act_in, src_seq=input_ids, bs=belief_state, input_mask=act_input_mask)
loss1 = ce_loss_func(
logits.contiguous().view(logits.size(0) * logits.size(1), -1).contiguous(),
act_out.contiguous().view(-1))
# response loss
resp_logits = resp_generator.resp_forward(tgt_seq=rep_in, src_seq=input_ids, act_vecs=act_vecs,
act_mask=action_masks, input_mask=resp_input_mask)
loss2 = ce_loss_func(
resp_logits.contiguous().view(resp_logits.size(0) * resp_logits.size(1), -1).contiguous(),
resp_out.contiguous().view(-1))
# overall loss
if epoch < 10:
loss = loss1
else:
loss = weighted_loss_func(loss1, loss2)
loss.backward()
optimizer.step()
if step % 100 == 0:
print("epoch {} \tstep {} training \ttotal_loss {:.6f} \tact_loss {:.6f} \tresp_loss {:.6f}".format(epoch, step, loss.item(), loss1.item(), loss2.item()))
alpha = min(1, alpha + 0.1 * epoch)
scheduler.step()
if loss2.item() < 3.0 and loss1.item() < 3.0 and epoch > 0 and epoch % args.evaluate_every == 0:
logger.info("start evaluating BLEU on validation set")
resp_generator.eval()
# Start Evaluating after each epoch
model_turns = {}
TP, TN, FN, FP = 0, 0, 0, 0
for batch_step, batch in enumerate(eval_dataloader):
all_pred = []
batch = tuple(t.to(device) for t in batch)
input_ids, action_masks, rep_in, resp_out, belief_state, \
bert_act_seq, act_in, act_out, all_label, act_input_mask, \
resp_input_mask, *_ = batch
if args.act_source == 'bert':
act_in = bert_act_seq
elif args.act_source == 'pred':
hyps, act_logits = resp_generator.act_translate_batch(input_mask=act_input_mask, bs=belief_state, \
src_seq=input_ids, n_bm=args.beam_size,
max_token_seq_len=Constants.ACT_MAX_LEN)
for hyp_step, hyp in enumerate(hyps):
pre1 = [0] * Constants.act_len
if len(hyp) < Constants.ACT_MAX_LEN:
hyps[hyp_step] = list(hyps[hyp_step]) + [Constants.PAD] * (Constants.ACT_MAX_LEN - len(hyp))
for w in hyp:
if w not in [Constants.PAD, Constants.EOS]:
pre1[w - 3] = 1
all_pred.append(pre1)
all_pred = torch.Tensor(all_pred)
all_label = all_label.cpu()
TP, TN, FN, FP = obtain_TP_TN_FN_FP(all_pred, all_label, TP, TN, FN, FP)
act_in = torch.tensor(hyps, dtype=torch.long).to(device)
else:
pass
_, _, act_vecs = resp_generator.act_forward(tgt_seq=act_in, src_seq=input_ids, bs=belief_state,
input_mask=act_input_mask)
action_masks = act_in.eq(Constants.PAD) + act_in.eq(Constants.EOS)
resp_hyps = resp_generator.resp_translate_batch(bs=belief_state, act_vecs=act_vecs,
act_mask=action_masks, input_mask=resp_input_mask,
src_seq=input_ids, n_bm=args.beam_size,
max_token_seq_len=40)
for hyp_step, hyp in enumerate(resp_hyps):
pred = tokenizer.convert_id_to_tokens(hyp)
file_name = val_id[batch_step * args.batch_size + hyp_step]
if file_name not in model_turns:
model_turns[file_name] = [pred]
else:
model_turns[file_name].append(pred)
precision = TP / (TP + FP + 0.001)
recall = TP / (TP + FN + 0.001)
F1 = 2 * precision * recall / (precision + recall + 0.001)
print("precision is {:.6f} recall is {:.6f} F1 is {:.6f}".format(precision, recall, F1))
logger.info("precision is {:.6f} recall is {:.6f} F1 is {:.6f}".format(precision, recall, F1))
BLEU = BLEU_calc.score(model_turns, gt_turns)
inform, request = evaluateModel(model_turns)
print("{} epoch, Validation BLEU {:.4f}, inform {:.2f}, request {:.2f}, score {:.2f}".format(epoch, BLEU, inform, request, (inform + request) / 2 + 100 * BLEU))
logger.info("{} epoch, Validation BLEU {:.4f}, inform {:.2f}, request {:.2f}, score {:.2f}".format(epoch, BLEU, inform, request, (inform + request) / 2 + 100 * BLEU))
if request > best_BLEU:
save_name = 'inform-{:.2f}-request-{:.2f}-bleu-{:.4f}-seed-{}'.format(inform, request, BLEU, args.seed)
torch.save(resp_generator.state_dict(), os.path.join(checkpoint_file, save_name))
best_BLEU = request
resp_file = os.path.join(args.output_file, 'resp_pred.json')
with open(resp_file, 'w') as fp:
model_turns = OrderedDict(sorted(model_turns.items()))
json.dump(model_turns, fp, indent=2)
resp_generator.train()
elif args.option == "test":
resp_generator.load_state_dict(torch.load(args.model))
logger.info("Loading model from {}".format(checkpoint_file))
resp_generator.eval()
# Start Evaluating after each epoch
model_turns = {}
act_turns={}
TP, TN, FN, FP = 0, 0, 0, 0
example_success={}
for batch_step, batch in enumerate(eval_dataloader):
all_pred = []
batch = tuple(t.to(device) for t in batch)
input_ids, action_masks, rep_in, resp_out, belief_state, bert_act_seq, act_in, act_out, all_label, \
act_input_mask, resp_input_mask, *_ = batch
if args.act_source == 'bert':
act_in = bert_act_seq
elif args.act_source == 'pred':
hyps, act_logits = resp_generator.act_translate_batch(input_mask=act_input_mask, bs=belief_state, \
src_seq=input_ids, n_bm=args.beam_size,
max_token_seq_len=Constants.ACT_MAX_LEN)
for hyp_step, hyp in enumerate(hyps):
pre1 = [0] * Constants.act_len
for w in hyp:
if w not in [Constants.PAD, Constants.EOS]:
pre1[w - 3] = 1
if len(hyp) < Constants.ACT_MAX_LEN:
hyps[hyp_step] = list(hyps[hyp_step]) + [Constants.PAD] * (Constants.ACT_MAX_LEN - len(hyp))
all_pred.append(pre1)
file_name = val_id[batch_step * args.batch_size + hyp_step]
if file_name not in act_turns:
act_turns[file_name] = [pre1]
else:
act_turns[file_name].append(pre1)
all_pred=torch.Tensor(all_pred)
all_label=all_label.cpu()
TP, TN, FN, FP = obtain_TP_TN_FN_FP(all_pred, all_label, TP, TN, FN, FP)
act_in = torch.tensor(hyps, dtype=torch.long).to(device)
else:
pass
_, _, act_vecs = resp_generator.act_forward(tgt_seq=act_in, src_seq=input_ids, bs=belief_state,
input_mask=act_input_mask)
action_masks = act_in.eq(Constants.PAD) + act_in.eq(Constants.EOS)
resp_hyps = resp_generator.resp_translate_batch(bs=belief_state, act_vecs=act_vecs, act_mask=action_masks,
input_mask=resp_input_mask,
src_seq=input_ids, n_bm=args.beam_size,
max_token_seq_len=40,gram_num=args.ngram)
for hyp_step, hyp in enumerate(resp_hyps):
pred = tokenizer.convert_id_to_tokens(hyp)
file_name = val_id[batch_step * args.batch_size + hyp_step]
if file_name not in model_turns:
model_turns[file_name] = [pred]
else:
model_turns[file_name].append(pred)
precision = TP / (TP + FP + 0.001)
recall = TP / (TP + FN + 0.001)
F1 = 2 * precision * recall / (precision + recall + 0.001)
print("precision is {:.6f} recall is {:.6f} F1 is {:.6f}".format(precision, recall, F1))
logger.info("precision is {:.6f} recall is {:.6f} F1 is {:.6f}".format(precision, recall, F1))
BLEU = BLEU_calc.score(model_turns, gt_turns)
inform, request = evaluateModel(model_turns, example_success)
print("Test BLEU {:.4f}, inform {:.2f}, request {:.2f}, score {:.2f}".format(BLEU, inform, request, (inform + request) / 2 + 100 * BLEU))
logger.info("Test BLEU {:.4f}, inform {:.2f}, request {:.2f}, score {:.2f}".format(BLEU, inform, request, (inform + request) / 2 + 100 * BLEU))
resp_file = os.path.join(args.output_file, 'resp_pred.json')
with open(resp_file, 'w') as fp:
model_turns = OrderedDict(sorted(model_turns.items()))
json.dump(model_turns, fp, indent=2)
act_file = os.path.join(args.output_file, 'act_pred.json')
with open(act_file, 'w') as fp:
act_turns = OrderedDict(sorted(act_turns.items()))
json.dump(act_turns, fp, indent=2)
with open('output/example_statistic.json','w') as f:
json.dump(example_success,f)
save_name = 'test-inform-{:.2f}-request-{:.2f}-bleu-{:.4f}'.format(inform, request, BLEU)
torch.save(resp_generator.state_dict(), os.path.join('model', save_name))
elif args.option == "postprocess":
resp_file = os.path.join(args.output_file, 'resp_pred.json')
with open(resp_file, 'r') as f:
model_turns = json.load(f)
success_rate = nondetokenize(model_turns, dialogs)
BLEU = BLEU_calc.score(model_turns, gt_turns)
print(BLEU)
resp_file = os.path.join(args.output_file, 'resp_non_delex_pred.json')
with open(resp_file, 'w') as fp:
model_turns = OrderedDict(sorted(model_turns.items()))
json.dump(model_turns, fp, indent=2)